openparser-sdk

Parse and extract structured data from documents with the OpenParser API.

Install the PyPI distribution openparser-sdk; import the client as openparser.

Documentation · OpenParser

Install

pip install openparser-sdk

Use Python 3.10 or newer. Create an API key in the OpenParser dashboard.

Quick Start

import os
from pathlib import Path
from openparser import OpenParserClient

client = OpenParserClient(api_key=os.environ["OPENPARSER_API_KEY"])

result = client.parse.sync(
    {"ocr_model": "paddleocr-vl-1.6", "output_format": "openparser@1"},
    file=Path("invoice.pdf"),
)

print(result.page_count)

Set OPENPARSER_API_KEY to create the client without constructor arguments. Set OPENPARSER_BASE_URL to use a different API origin.

Parse

The API holds synchronous requests for up to 300 seconds. It returns the result when processing finishes or a durable job reference when the wait expires.

# Sync: hold the connection until the result or timeout.
parsed = client.parse.sync(
    {"ocr_model": "paddleocr-vl-1.6"},
    file=Path("document.pdf"),
)

# Async: create a job and poll it later.
accepted = client.parse.async_(
    {"ocr_model": "paddleocr-vl-1.6"},
    file=Path("document.pdf"),
)
job = client.wait_for_job(accepted.id)

# Reuse a file-pool upload instead of inline bytes.
uploaded = client.files.upload(Path("document.pdf"))
parsed = client.parse.sync({"ocr_model": "paddleocr-vl-1.6", "file_id": uploaded.id})

The SDK adds an Idempotency-Key header to every parse and extract request. Pass idempotency_key= when you need to control retries.

Extract

extracted = client.extract.sync(
    {
        "ocr_model": "paddleocr-vl-1.6",
        "llm_model": "openai/gpt-6-luna",
        "schema": {
            "type": "object",
            "properties": {"total": {"type": "number"}},
        },
    },
    file=Path("invoice.pdf"),
)

suggested = client.extract.suggest_schema(
    {
        "parse_job_id": "opj_...",
        "hint": "Invoice number, vendor, and total",
    }
)

Grounding and lineage

Set "grounding": "field" to receive verified citations and a lineage@1 derivation DAG:

extracted = client.extract.sync(
    {
        "ocr_model": "mistral-ocr-4",
        "llm_model": "openai/gpt-6-luna",
        "grounding": "field",
        "schema": {
            "type": "object",
            "properties": {"total": {"type": "number"}},
        },
    },
    file=Path("invoice.pdf"),
)

lineage = extracted.to_dict().get("lineage")
if lineage is not None:
    print(lineage["outputs"])

The graph connects each output value to its source evidence and the operations that produced it. Evidence carries the closest recognition confidence supplied by the selected OCR model. Applications can append normalization, calculation, inference, and human-review activities without replacing the original machine result.

OpenParserClient is synchronous. Methods named async_ submit durable jobs without waiting for processing to finish.

Jobs

jobs_page = client.jobs.list(status="succeeded", limit=25)
first_job_id = jobs_page.data[0].id
job = client.jobs.get("opj_...")
parse_result = client.jobs.result("opj_...", format="openparser@1")
source_bytes = client.jobs.source("opj_...")

jobs.result() returns parse representations. Extract output is available on the job returned by jobs.get().

Files

uploaded = client.files.upload(Path("contract.pdf"))
metadata = client.files.get(uploaded.id)
content = client.files.download(uploaded.id)
client.files.delete(uploaded.id)

Uploads accept pathlib.Path, a file handle, or {"content": bytes, "filename": str, "mime_type": str?}.

Models

ocr_models = client.models.list_ocr()
llm_models = client.models.list_llm(mode="search", q="claude")
first_ocr_model = ocr_models.data[0]
first_llm_model = llm_models.data[0]

Pipelines

pipeline = client.pipelines.create(
    {
        "name": "invoice-extract",
        "ocr_model": "paddleocr-vl-1.6",
        "llm_model": "anthropic/claude-sonnet-5.5",
        "schema": {
            "type": "object",
            "properties": {"vendor": {"type": "string"}},
            "required": ["vendor"],
        },
    }
)

listed = client.pipelines.list()
first_pipeline = listed.items[0]
current = client.pipelines.get(pipeline.id)
updated = client.pipelines.update(pipeline.id, {"name": "invoice-v2"})
client.pipelines.delete(pipeline.id)

Errors

Every non-2xx response raises a typed subclass of OpenParserError:

Each error exposes the API response through error.envelope: code, message, request_id, and retryable.

Development

After changing the OpenAPI specification, regenerate the client and run its checks:

packages/openparser-sdk-python/scripts/codegen.sh
packages/openparser-sdk-python/scripts/check-codegen.sh
uv run --project packages/openparser-sdk-python pytest

License

Apache-2.0